Modern Smart City AI systems are empowered by various smart sensors. For example, traffic monitor sensors, snow/rain detectors, LIDAR on self-driving cars, etc. In these scenarios, we often encounter the following challenges:
1) Real-time data collection. The volume of the data is too large to upload to the center super-computer in real-time.
2) High requirements on the robustness and power-consumption. The system is often driven by battery. It must be able to work without external power supply for several hours.
3) Very limited memory.
In these scenarios, the conventional deep learning models cannot work well, due to their huge power consumption and memory occupation. Therefore, it is important to customize novel deep learning models for these scenarios.
- Innovate a new deep architecture optimized for battery-powered devices.
- Develop a deep learning pipeline for searching and training deep networks.
- Verify system performance on public benchmark datasets, including but not limited to FLOPs, power wattage, inference speed.
- Using MobileNet as baseline, achieve the same top-1 accuracies on ImageNet-1k with 30% less FLOPs.
- On MCU chip, accelerate the inference speed 20% v.s. MobileNet.
- On MCU chip, consume 20% less battery power v.s. MobileNet.
- Publish 1~2 CCF-A papers.
Related Research Topics
- Deep Network pruning and quantization
- Lightweight vision transformers
- Neural Architecture Search for TinyML
- Latency and power-consumption prediction for edge devices